Start a project

Independent AI studio

AI that plays,decides anddreams.

We build intelligent systems for games, engineering teams and asset production. Research-grade models, engineered to survive production.

Start a projectSee what we build
  1. 01PlayAI games & living worlds
  2. 02DecideAutonomous triage & remediation
  3. 03DreamGenerative asset pipelines

Disciplines

Three kinds of intelligence.One studio.

We work where research meets production. Each discipline below comes with a small simulation running live in your browser. Go ahead and poke at them.

01 play

AI games & living worlds

Characters that remember, adapt and surprise, plus the assets to fill their worlds. We design NPC behaviour, dialogue systems, procedural content and asset generation for games.

  • NPC behaviour, memory & dialogue models
  • Procedural levels, quests and encounters
  • Game asset generation: items, icons, textures
  • Player modelling & adaptive difficulty
Specimen 01 · Emergent NPCsLive simulation

Interactive simulation: a crowd of small NPCs that drift towards the pointer when it moves slowly and scatter when it moves fast. Each NPC remembers how it was treated.

Move slowly near them to earn trust. Move fast and they scatter, and they remember.Trust 0.30Curious 0Wary 0

02 decide

Autonomous triage & remediation

Not a smarter inbox. An end-to-end fixer. It understands user requests, feedback and bug reports, traces each one to its root cause, then writes, tests and ships the fix. Fully automated, shaped around the workflow your engineers already use.

  • Understands requests, feedback, bugs & alerts
  • Root-cause analysis across logs, traces and code
  • Writes, tests and ships the fix automatically
  • Your workflow: your stages, gates and rules
Specimen 02 · Report → fix pipelineLive simulation

Bug14 user reports · 1 alertCase #2481

Checkout fails for EU customers since 09:14

  1. IntakeMerged 14 reports and 1 alert into a single case
  2. UnderstandBug · P1 · checkout · only EU locales affected
  3. DiagnoseRoot cause: amount formatter throws on a missing locale, introduced in a3f9c1e (deployed 09:11)
  4. FixPR #482: locale fallback + regression test (+12 −3)
  5. VerifyCI green · canary at 5% · error rate 4.1% → 0.02%
  6. ShipRolled out to 100% · all 14 reporters notified
Fully automated
Simulated cases. Flip the switch to add a human approval gate. The workflow is yours to shape.Cases 1Shipped 1Escalated 0

03 dream

Generative asset pipelines

From pure noise to production-ready assets. We build diffusion pipelines that design, render, de-duplicate and catalogue at scale: textiles and garments, game art, product imagery.

  • Textile, print & garment generation
  • Game assets, icons & concept art
  • Dedup, review & catalogue publishing built in
  • Custom fine-tunes and cost-aware GPU serving
Specimen 03 · Reverse diffusionLive simulation

Promptseamless indigo block-print textile, hand-stamped white floral motifs on natural linen

Pick an asset, then drag the slider to scrub through the sampler.Step 50/50ᾱ 1.000Seed 1337

Why we exist

Most AI never leaves the demo. We build the kind that ships: models that run on budget, systems that fix what they find, and worlds that feel alive after the lights go out.

Selected work

Built after dark.

A few things we've made and are making: games, internal tools and pipelines that run quietly in production.

Featured caseAsset generation · Textiles & garments2026

A catalogue that designs itself overnight.

An unattended pipeline behind a live product catalogue. Every night it designs new garments and prints, renders them on GPU, weeds out near-duplicates, previews each piece as worn and hands a reviewer a deck that is ready to approve.

  • Variety is dealt, not hoped for: each batch draws its design briefs without replacement, so rare styles actually show up.
  • Model grammar lives in config. Swapping diffusion checkpoints changes the prompts, not the generator.
  • Self-healing: a sweeper recovers stuck renders and orphaned vectors, and degraded runs report themselves loudly.
  • Structured LLM output
  • SDXL
  • OpenCLIP
  • Vector search
  • Go
  • Python
Nightly batch · review deckClearFlaggedBlocked

  1. 01 Block-print floralClear

  2. 02 Madder ikatClear

  3. 03 Sage tartanClear

  4. 04 Plum paisleyClear

  5. 05 Indigo floral IIFlagged ≈ 01

  6. 06 Ochre herringboneClear

  7. 07 Gold latticeClear

  8. 08 Madder ikatBlocked = 02

Illustrative batch. Near-duplicates are caught twice: perceptual hash and embedding distance.

  1. 01Design briefBriefs are dealt across garment family, pattern, colour, fit and framing, so variety is built in.
  2. 02LLM designA language model writes each piece as structured facts: material, construction, details.
  3. 03Prompt composerFacts become model-specific prompts from per-checkpoint config.
  4. 04GPU renderDiffusion workers render a catalogue shot and an icon for every piece.
  5. 05DedupPerceptual hash + embedding distance sort renders into clear, flagged or blocked.
  6. 06Worn previewEach piece is also rendered as worn, so fit gets judged, not just a flat shot.
  7. 07Review & publishA human approves; publishing into the catalogue is transactional.

  • Triage & remediation2026

    Report-to-fix automation

    An end-to-end system that reads user requests, feedback and bug reports, finds the root cause and ships a tested fix, all inside the engineering team’s own workflow.

    • Root-cause analysis
    • Auto-remediation
    • Custom workflows
    Case study soon

  • AI game2026

    Lantern

    A narrative exploration game where every resident of a sleeping town is driven by a memory-aware dialogue model.

    • Dialogue models
    • NPC memory
    • Game design
    Case study soon

  • ResearchOngoing

    Night Shift

    Ongoing R&D into small, fast models that can reason for game characters on-device, without a round-trip to the cloud.

    • Small models
    • On-device
    • Distillation
    Case study soon

Capabilities

From the first notebook to the pager.

A small senior team covering the full lifecycle: data, modelling, serving and the product around it. No hand-offs between “research” and “engineering”. It’s the same people.

  1. 01

    Model fine-tuning

    LLMs, diffusion models and classifiers adapted to your data, voice and constraints.

  2. 02

    Inference & serving

    Quantisation, batching and GPU scheduling so models stay fast and affordable in production.

  3. 03

    Triage & remediation

    Report-to-fix automation: understanding, root-cause analysis and tested fixes, wired into your engineering workflow.

  4. 04

    Game AI

    NPC behaviour, dialogue, procedural generation and player modelling for game teams of any size.

  5. 05

    Asset generation

    Textiles, garments, game art and product imagery, generated, de-duplicated and reviewed at catalogue scale.

  6. 06

    Evaluation & safety

    Evals, red-teaming and monitoring that tell you when a model drifts, before your users do.

Tools we work with

  • PyTorch
  • Diffusers
  • Transformers
  • vLLM
  • SDXL
  • Flux
  • LoRA
  • ONNX
  • Python
  • Go
  • TypeScript
  • Redis
  • Qdrant
  • OpenCLIP
  • PostgreSQL
  • Modal
  • Cloudflare
  • Unity
  • Godot

Process

From signal to shipped.

Short loops, working software early, and honest numbers at every step. You always know what the model can do, and what it can't do yet.

  1. 01Week 1

    Signal

    We map the problem, the data you have and what “good” looks like. You leave with a scoped plan and an honest feasibility call.

  2. 02Weeks 2-4

    Prototype

    A working model on your real data, wrapped in something you can click. Early enough to change direction cheaply.

  3. 03Weeks 4-8

    Harden

    Evals, latency and cost budgets, failure modes and safety checks. The unglamorous part that makes it shippable.

  4. 04Ongoing

    Operate

    Deployment, monitoring and retraining. Run by us, or handed over to your team with full documentation.

Contact

Bring us the problem that keeps you up at night.

A game that needs characters with a mind of their own, a bug backlog that never stops growing, a catalogue that needs a thousand new assets by Monday. Tell us about it. The first call is free and comes with an honest opinion.

Studio time
--:-- UTC
Based
Remote-first
Status
Taking new projects